Reproduce “Multi-Agent Computer Use” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried
MACU coordinates computer-use tasks by dividing them into subtasks that workers can execute in parallel. A manager tracks a task graph, sends ready subtasks to workers, and revises work with new findings, passing findings and state between partially observable environments. On Odysseys, success rose from 8.5% to 34.0%, and median time fell from 162.4 to 110.3 minutes, but not on every benchmark.
Key insights
- Orchestrating homogeneous CUA subagents with a dynamic DAG-based manager is effective for complex, long-horizon computer interaction tasks.
- Continuous replanning and explicit state transfer via file management are crucial for managing partial observability and improving task success in dynamic computer environments.
- Multi-agent systems can scale more effectively with increased inference budget, leveraging more compute to solve tasks beyond single-agent capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.